When Is Operation Ordering Required in Replicated Transactional Storage?

Total Page:16

File Type:pdf, Size:1020Kb

When Is Operation Ordering Required in Replicated Transactional Storage? When Is Operation Ordering Required in Replicated Transactional Storage? Irene Zhang, Naveen Kr. Sharma, Adriana Szekeres, Arvind Krishnamurthy, and Dan R. K. Ports University of Washington iyzhang,naveenks,aaasz,arvind,drkp @cs.washington.edu { } Abstract Today’s replicated transactional storage systems typically have a layered architecture, combining protocols for transaction coordination, consistent replication, and concurrency control. These systems generally require costly strongly-consistent replication protocols like Paxos, which assign a total order to all operations. To avoid this cost, we ask whether all replicated operations in these systems need to be strictly ordered. Recent research has yielded replication protocols that can avoid unnecessary ordering, e.g., by exploiting commutative operations, but it is not clear how to apply these to replicated transaction processing systems. We answer this question by analyzing existing transaction processing designs in terms of which replicated operations require ordering and which simply require fault tolerance. We describe how this analysis leads to our recent work on TAPIR, a transaction protocol that efficiently provides strict serializability by using a new replication protocol that provides fault tolerance but not ordering for most operations. 1 Introduction Distributed storage systems for today’s large-scale web applications must meet a daunting set of requirements: they must offer high performance, graceful scalability, and continuous availability despite the inevitablity of failures. Increasingly, too, application programmers prefer systems that support distributed transactions with strong consistency to help them manage application complexity and concurrency in a distributed environment. Several recent systems [11, 3, 8, 5, 6] reflect this trend. One notable example is Google’s Spanner system [6], which guarantees strictly-serializable (aka linearizable or externally consistent) transaction ordering [6]. Generally, distributed transactional storage with strong consistency comes at a heavy performance price. These systems typically integrate several expensive mechanisms, including concurrency control schemes like strict two-phase locking, strongly consistent replication protocols like Paxos, and atomic commitment protocols like two-phase commit. The costs associated with these mechanisms often drive application developers to use more efficient, weak consistency protocols that fail to provide strong system guarantees. In conventional designs, these protocols – concurrency control, replication, and atomic commitment – are implemented in separate layers, with each layer providing a subset of the guarantees required for distributed Copyright 2016 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE. Bulletin of the IEEE Computer Society Technical Committee on Data Engineering 27 ACID transactions. For example, a system may partition data into shards, where each shard is replicated using Paxos, and use two-phase commit and two-phase locking to implement serializable transactions across shards. Though now frequently implemented together, these protocols were originally developed to address separate issues in distributed systems. In our recent work [23, 24], we have taken a different approach: we consider the storage system architecture as a whole and identify opportunities for cross-layer optimizations to improve performance. To improve the performance of replicated transactional systems, we ask the question: do all replicated op- erations need to be strictly ordered? Existing systems treat the replication protocol as an implementation of a persistent, ordered log. Ensuring a total ordering, however, require cross-replica coordination on every oper- ation, increasing system latency, and is typically implemented with a designated leader, introducing a system bottleneck. A recent line of distributed systems research has developed more efficient replication protocols by identifying cases when operations do not need to be ordered – typically by having the developer express which operations commute with others [4, 14, 18, 15]. Yet it is not obvious how to apply these techniques to the lay- ered design of transactional storage systems: the operations being replicated are not transactions themselves, but coordination operations like prepare or commit records. To answer this question, we analyzed the interaction of atomic commitment, consistent replication, and concurrency control protocols in common combinations. We consider both their requirements on layers below and the guarantees they provide for layers above. Our analysis leads to several key insights about the protocols implemented in distributed transactional storage systems: • The concurrency control mechanism is the application for the replication layer, so its requirements impact the replication algorithm. • Unlike disk-based durable storage, replication can separate ordering/consistency and fault-tolerance guar- antees with different costs for each. • Consensus-based replication algorithms like Paxos are not the most efficient way to provide ordering. • Enforcing consistency at every layer is not necessary for maintaining global transactional consistency. We describe how these observations led us to design a new transactional storage system, TAPIR [23, 24]. TAPIR provides strict serializable isolation of transactions, but relies on a weakly consistent underlying repli- cation protocol, inconsistent replication (IR). IR is designed to efficiently support fault-tolerant but unordered operations, and TAPIR uses an optimistic timestamp ordering technique to make most of its replicated oper- ations unordered. This article does not attempt to provide a complete explanation of the TAPIR protocol, its performance, or its correctness. Rather, it analyzes it along with additional protocols to demonstrate the value of separating operation ordering from fault tolerance in replicated transactional systems. 2 Coordination Protocols This paper is about the protocols used to build scalable, fault-tolerant distributed storage system, like distributed key-value stores or distributed databases. In this section, we specify the category of systems we are considering, and review the standard protocols that they use to coordinate replicated, distributed transactions. 2.1 Transactional System Model The protocols that we discuss in this paper are designed for scalable, distributed transactional storage. We as- sume the storage system divides stored data across several shards. Within a shard, we assume the system uses replication for availability and fault tolerance. Each shard is replicated across a set of storage servers organized 28 2PC Distributed Transaction CC CC CC Protocol Replication R R R R R R R R R Protocol Figure 1: Standard protocol architecture for transactional distributed storage systems. into a replica group. Each replica group consists of 2 f + 1 storage servers, where f is the number of faults that can be tolerated by one replica group. Transactions may read and modify data that spans multiple shards; a distributed transaction protocol ensures that transactions are applied consistently to all shards. We consider system architectures that layer the distributed transaction protocol atop the replication protocol, as shown in Figure 1. In the context of Agrawal et al.’s recent taxonomy of partitioned replicated databases [2], these are replicated object systems. This is the traditional system architecture, used by several influential sys- tems including Spanner [6], MDCC [11], Scatter [9] and Granola [7]. However, other architectures have been proposed that run the distributed transaction protocol within each site, and use replication across sites [17]. Our analysis does not consider the cost of disk writes. This may seem an unconventional choice, as syn- chronous writes to disk are the traditional way to achieve durability in storage systems – and a major source of their overhead. Rather, we consider architectures where durability is achieved using in-memory replication, combined with asynchronous checkpoints to disk. Several recent systems (e.g., RAMCloud [20], H-Store [21]) have adopted this approach, which offers two advantages over disk. First, it ensures that the data remains contin- uously available when one system fails, minimizing downtime. Second, recording an operation in memory on multiple machines can have far lower latency than synchronously writing to disk, while still tolerating common failures. We assume that clients are application servers that access data stored in the system. Clients have access to a directory of storage servers and are able to directly map data to storage servers, using a technique like consistent hashing [10]. Transaction Model We assume a general transaction model. Clients begin a transaction, then execute oper- ations during the transaction’s execution period. During this period, the client is free to abort the transaction. Once the client finishes execution, it can commit the transaction, atomically and durably committing the exe- cuted operations to the storage servers. 2.2 Standard Protocols The standard architecture for these distributed storage systems, shown in Figure 1 is layering an atomic com- mitment protocol, like
Recommended publications
  • ACID Compliant Distributed Key-Value Store
    ACID Compliant Distributed Key-Value Store #1 #2 #3 Lakshmi Narasimhan Seshan ,​ Rajesh Jalisatgi ,​ Vijaeendra Simha G A # ​ ​ ​ ​1l​ [email protected] # ​2r​ [email protected] #3v​ [email protected] Abstract thought that would be easier to implement. All Built a fault-tolerant and strongly-consistent key/value storage components run http and grpc endpoints. Http endpoints service using the existing RAFT implementation. Add ACID are for debugging/configuration. Grpc Endpoints are transaction capability to the service using a 2PC variant. Build used for communication between components. on the raftexample[1] (available as part of etcd) to add atomic Replica Manager (RM): RM is the service discovery transactions. This supports sharding of keys and concurrent transactions on sharded KV Store. By Implementing a part of the system. RM is initiated with each of the transactional distributed KV store, we gain working knowledge Replica Servers available in the system. Users have to of different protocols and complexities to make it work together. instantiate the RM with the number of shards that the user wants the key to be split into. Currently we cannot Keywords— ACID, Key-Value, 2PC, sharding, 2PL Transaction dynamically change while RM is running. New Replica ​ ​ Servers cannot be added or removed from once RM is I. INTRODUCTION ​ up and running. However Replica Servers can go down Distributed KV stores have become a norm with the and come back and RM can detect this. Each Shard advent of microservices architecture and the NoSql Leader updates the information to RM, while the DTM revolution. Initially KV Store discarded ACID properties leader updates its information to the RM.
    [Show full text]
  • Lecture 15: March 28 15.1 Overview 15.2 Transactions
    CMPSCI 677 Distributed and Operating Systems Spring 2019 Lecture 15: March 28 Lecturer: Prashant Shenoy 15.1 Overview This section covers the following topics: Distributed Transactions: ACID Properties, Concurrency Control, Serializability, Two-Phase Locking 15.2 Transactions Transactions are used widely in databases as they provide atomicity (all or nothing property). Atomicity is the property in which operations inside a transaction all execute successfully or none of the operations execute. Thus, if a process crashes during a transaction, all operations that happened during that transaction must be undone. Transactions are usually implemented using locks. Without locks there would be interleaving of operations from two transactions, causing overwrites and inconsistencies. It has to look like all instructions/operations of one transaction are executed at once, while ensuring other external instructions/operations are not executing at the same time. This is especially important in real-world applications such as banking applications. 15.2.1 ACID: Most transaction systems will provide the following properties • Atomicity: All or nothing (all instructions execute or none; cannot have some instructions execute and some do not) • Consistency: Transaction takes system from one consistent (valid) state to another • Isolated: Transaction executes as if it is the only transaction in the system. Ensures that the concurrent execution of transactions results in a system state that would be obtained if transactions were executed sequentially (serializability).
    [Show full text]
  • Distributed Transactions
    Distributed Systems 600.437 Distributed Transactions Department of Computer Science The Johns Hopkins University Yair Amir Fall 16/ Lecture 6 1 Distributed Transactions Lecture 6 Further reading: Concurrency Control and Recovery in Database Systems P. A. Bernstein, V. Hadzilacos, and N. Goodman, Addison Wesley. 1987. Transaction Processing: Concepts and Techniques Jim Gray & Andreas Reuter, Morgan Kaufmann Publishers, 1993. Yair Amir Fall 16/ Lecture 6 2 Transaction Processing System Clients Messages to TPS outside world Real actions (firing a missile) Database Yair Amir Fall 16/ Lecture 6 3 Basic Definition Transaction - a collection of operations on the physical and abstract application state, with the following properties: • Atomicity. • Consistency. • Isolation. • Durability. The ACID properties of a transaction. Yair Amir Fall 16/ Lecture 6 4 Atomicity Changes to the state are atomic: - A jump from the initial state to the result state without any observable intermediate state. - All or nothing ( Commit / Abort ) semantics. - Changes include: - Database changes. - Messages to outside world. - Actions on transducers. (testable / untestable) Yair Amir Fall 16/ Lecture 6 5 Consistency - The transaction is a correct transformation of the state. This means that the transaction is a correct program. Yair Amir Fall 16/ Lecture 6 6 Isolation Even though transactions execute concurrently, it appears to the outside observer as if they execute in some serial order. Isolation is required to guarantee consistent input, which is needed for a consistent program to provide consistent output. Yair Amir Fall 16/ Lecture 6 7 Durability - Once a transaction completes successfully (commits), its changes to the state survive failures (what is the failure model ? ).
    [Show full text]
  • A Study of the Availability and Serializability in a Distributed Database System
    A STUDY OF THE AVAILABILITY AND SERIALIZABILITY IN A DISTRIBUTED DATABASE SYSTEM David Wai-Lok Cheung B.Sc., Chinese University of Hong Kong, 1971 M.Sc., Simon Fraser University, 1985 A THESIS SUBMI'ITED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHLLOSOPHY , in the School of Computing Science 0 David Wai-Lok Cheung 1988 SIMON FRASER UNIVERSITY January 1988 All rights reserved. This thesis may not be reproduced in whole or in part, by photocopy or other means, without permission of the author. 1 APPROVAL Name: David Wai-Lok Cheung Degree: Doctor of Philosophy Title of Thesis: A Study of the Availability and Serializability in a Distributed Database System Examining Committee: Chairperson: Dr. Binay Bhattacharya Senior Supervisor: Dr. TikoJameda WJ ru p v Dr. Arthur Lee Liestman Dr. Wo-Shun Luk Dr. Jia-Wei Han External Examiner: Toshihide Ibaraki Department of Applied Mathematics and Physics Kyoto University, Japan Bate Approved: January 15, 1988 PARTIAL COPYRIGHT LICENSE I hereby grant to Simon Fraser University the right to lend my thesis, project or extended essay (the title of which is shown below) to users of the Simon Fraser University Library, and to make partial or single copies only for such users or in response to a request from the library of any other university, or other educational institution, on its own behalf or for one of its users. I further agree that permission for multiple copying of this work for scholarly purposes may be granted by me or the Dean of Graduate Studies. It is understood that copying or publication of this work for financial gain shall not be allowed without my written permission, T it l e of Thes i s/Project/Extended Essay Author: (signature) ( name (date) ABSTRACT Replication of data objects enhances the reliability and availability of a distributed database system.
    [Show full text]
  • An Alternative Model to Overcoming Two Phase Commit Blocking Problem
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by International Journal of Computer (IJC - Global Society of Scientific Research and... International Journal of Computer (IJC) ISSN 2307-4523 (Print & Online) © Global Society of Scientific Research and Researchers http://ijcjournal.org/ An Alternative Model to Overcoming Two Phase Commit Blocking Problem Hassan Jafar Sheikh Salaha*, Dr. Richard M. Rimirub , Dr. Michael W. Kimwelec a,b,cComputer Science, Jomo Kenyatta University of Agriculture and Technology Kenya aEmail: [email protected] bEmail: [email protected] cEmail: [email protected] Abstract In distributed transactions, the atomicity requirement of the atomic commitment protocol should be preserved. The two phased commit protocol algorithm is widely used to ensure that transactions in distributed environment are atomic. However, a main drawback was attributed to the algorithm, which is a blocking problem. The system will get in stuck when participating sites or the coordinator itself crashes. To address this problem a number of algorithms related to 2PC protocol were proposed such as back up coordinator and a technique whereby the algorithm passes through 3PC during failure. However, both algorithms face limitations such as multiple site and backup coordinator failures. Therefore, we proposed an alternative model to overcoming the blocking problem in combined form. The algorithm was simulated using Britonix transaction manager (BTM) using Eclipse IDE and MYSQL. In this paper, we assessed the performance of the alternative model and found that this algorithm handled site and coordinator failures in distributed transactions using hybridization (combination of two algorithms) with minimal communication messages.
    [Show full text]
  • Transaction Management in the R* Distributed Database Management System
    Transaction Management in the R* Distributed Database Management System C. MOHAN, B. LINDSAY, and R. OBERMARCK IBM Almaden Research Center This paper deals with the transaction management aspects of the R* distributed database system. It concentrates primarily on the description of the R* commit protocols, Presumed Abort (PA) and Presumed Commit (PC). PA and PC are extensions of the well-known, two-phase (2P) commit protocol. PA is optimized for read-only transactions and a class of multisite update transactions, and PC is optimized for other classes of multisite update transactions. The optimizations result in reduced intersite message traffic and log writes, and, consequently, a better response time. The paper also discusses R*‘s approach toward distributed deadlock detection and resolution. Categories and Subject Descriptors: C.2.4 [Computer-Communication Networks]: Distributed Systems-distributed datahes; D.4.1 [Operating Systems]: Process Management-concurrency; deadlocks, syndvonization; D.4.7 [Operating Systems]: Organization and Design-distributed sys- tems; D.4.5 [Operating Systems]: Reliability--fault tolerance; H.2.0 [Database Management]: General-concurrency control; H.2.2 [Database Management]: ‘Physical Design-recouery and restart; H.2.4 [Database Management]: Systems-ditributed systems; transactionprocessing; H.2.7 [Database Management]: Database Administration-logging and recouery General Terms: Algorithms, Design, Reliability Additional Key Words and Phrases: Commit protocols, deadlock victim selection 1. INTRODUCTION R* is an experimental, distributed database management system (DDBMS) developed and operational at the IBM San Jose Research Laboratory (now renamed the IBM Almaden Research Center) 118, 201. In a distributed database system, the actions of a transaction (an atomic unit of consistency and recovery [13]) may occur at more than one site.
    [Show full text]
  • Implementing Distributed Transactions Distributed Transaction Distributed Database Systems ACID Properties Global Atomicity Atom
    Distributed Transaction • A distributed transaction accesses resource managers distributed across a network Implementing Distributed • When resource managers are DBMSs we refer to the Transactions system as a distributed database system Chapter 24 DBMS at Site 1 Application Program DBMS 1 at Site 2 2 Distributed Database Systems ACID Properties • Each local DBMS might export • Each local DBMS – stored procedures, or – supports ACID properties locally for each subtransaction – an SQL interface. • Just like any other transaction that executes there • In either case, operations at each site are grouped – eliminates local deadlocks together as a subtransaction and the site is referred • The additional issues are: to as a cohort of the distributed transaction – Global atomicity: all cohorts must abort or all commit – Each subtransaction is treated as a transaction at its site – Global deadlocks: there must be no deadlocks involving • Coordinator module (part of TP monitor) supports multiple sites ACID properties of distributed transaction – Global serialization: distributed transaction must be globally serializable – Transaction manager acts as coordinator 3 4 Atomic Commit Protocol Global Atomicity Transaction (3) xa_reg • All subtransactions of a distributed transaction Manager Resource must commit or all must abort (coordinator) Manager (1) tx_begin (cohort) • An atomic commit protocol, initiated by a (4) tx_commit (5) atomic coordinator (e.g., the transaction manager), commit protocol ensures this. (3) xa_reg Resource Application – Coordinator
    [Show full text]
  • Distributed Transaction Processing: Reference Model, Version 3
    Guide Distributed Transaction Processing: Reference Model, Version 3 HNICA C L E G T U I D E S [This page intentionally left blank] X/Open Guide Distributed Transaction Processing: Reference Model, Version 3 X/Open Company Ltd. February 1996, X/Open Company Limited All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission of the copyright owners. X/Open Guide Distributed Transaction Processing: Reference Model, Version 3 ISBN: 1-85912-170-5 X/Open Document Number: G504 Published by X/Open Company Ltd., U.K. Any comments relating to the material contained in this document may be submitted to X/Open at: X/Open Company Limited Apex Plaza Forbury Road Reading Berkshire, RG1 1AX United Kingdom or by Electronic Mail to: [email protected] ii X/Open Guide Contents Chapter 1 Introduction............................................................................................... 1 1.1 Overview ...................................................................................................... 1 1.2 Benefits of X/Open DTP ........................................................................... 1 1.3 Areas Not Addressed................................................................................. 2 1.4 Relationship to International Standards................................................ 2 Chapter 2 Definitions.................................................................................................
    [Show full text]
  • Fast General Distributed Transactions with Opacity
    Fast General Distributed Transactions with Opacity Alex Shamis∗ Matthew Renzelmann Stanko Novakovic† Microsoft Research Microsoft VMware Georgios Chatzopoulos‡ Aleksandar Dragojević Dushyanth Narayanan EPFL Microsoft Research Microsoft Research Miguel Castro Microsoft Research ABSTRACT KEYWORDS Transactions can simplify distributed applications by hid- distributed transactions, RDMA, opacity, global time, clock ing data distribution, concurrency, and failures from the synchronization, multi-version concurrency control application developer. Ideally the developer would see the ACM Reference Format: abstraction of a single large machine that runs transactions Alex Shamis, Matthew Renzelmann, Stanko Novakovic, Geor- sequentially and never fails. This requires the transactional gios Chatzopoulos, Aleksandar Dragojević, Dushyanth Narayanan, subsystem to provide opacity (strict serializability for both and Miguel Castro. 2019. Fast General Distributed Transactions committed and aborted transactions), as well as transparent with Opacity. In 2019 International Conference on Management of fault tolerance with high availability. As even the best ab- Data (SIGMOD ’19), June 30-July 5, 2019, Amsterdam, Netherlands. stractions are unlikely to be used if they perform poorly, the ACM, New York, NY, USA, 16 pages. https://doi.org/10.1145/3299869. system must also provide high performance. 3300069 Existing distributed transactional designs either weaken this abstraction or are not designed for the best performance 1 INTRODUCTION within a data center. This paper extends the design of FaRM Cloud data centers provide many relatively small, individu- — which provides strict serializability only for committed ally unreliable servers. Cloud services need to run on clus- transactions — to provide opacity while maintaining FaRM’s ters of such servers to maintain availability despite indi- high throughput, low latency, and high availability within vidual server failures.
    [Show full text]
  • Salt: Combining ACID and BASE in a Distributed Database
    Salt: Combining ACID and BASE in a Distributed Database Chao Xie, Chunzhi Su, Manos Kapritsos, Yang Wang, Navid Yaghmazadeh, Lorenzo Alvisi, and Prince Mahajan, The University of Texas at Austin https://www.usenix.org/conference/osdi14/technical-sessions/presentation/xie This paper is included in the Proceedings of the 11th USENIX Symposium on Operating Systems Design and Implementation. October 6–8, 2014 • Broomfield, CO 978-1-931971-16-4 Open access to the Proceedings of the 11th USENIX Symposium on Operating Systems Design and Implementation is sponsored by USENIX. Salt: Combining ACID and BASE in a Distributed Database Chao Xie, Chunzhi Su, Manos Kapritsos, Yang Wang, Navid Yaghmazadeh, Lorenzo Alvisi, Prince Mahajan The University of Texas at Austin Abstract: This paper presents Salt, a distributed recently popularized by several NoSQL systems [1, 15, database that allows developers to improve the perfor- 20, 21, 27, 34]. Unlike ACID, BASE offers more of a set mance and scalability of their ACID applications through of programming guidelines (such as the use of parti- the incremental adoption of the BASE approach. Salt’s tion local transactions [32, 37]) than a set of rigorously motivation is rooted in the Pareto principle: for many ap- specified properties and its instantiations take a vari- plications, the transactions that actually test the perfor- ety of application-specific forms. Common among them, mance limits of ACID are few. To leverage this insight, however, is a programming style that avoids distributed Salt introduces BASE transactions, a new abstraction transactions to eliminate the performance and availabil- that encapsulates the workflow of performance-critical ity costs of the associated distributed commit protocol.
    [Show full text]
  • Concurrency Control Protocol in Distributed System
    Concurrency Control Protocol In Distributed System Sonsie Burke humours: he phenomenalized his fords blinking and movelessly. Caroline Vic devocalize: he supinating his falterings sicker and ungrammatically. Ramsey is mercantilism and gnarl earthward while unafraid Guthrey deny and dozes. The message from similar as an approach certain has to distributed concurrency in system keeps selecting this method that Alert acm transactions enter creating more importantly, distributed database equivalent operating system, distributed in commit or more specifying tablespaces arranged optimization under high data replicated databases or prevents other. This approach is controlling concurrent transactions. It last this by sending the candidate an INQUIRE. If all systems, control protocol does not part is controlling access controls such messages and concurrently. ACP if significant site failures, if any. For distributed system is calculating an item would be acid properties as checking for detecting failures and control protocols. Aq when in concurrency control permits an increase in fig. Suppose a, a sink, that will be times where target database is inconsistent. MVCC for transaction support. Find answer your specific questions by searching them here. The concurrency in that message to guarantee serialisability are taken by controlling concurrent threads. We determine whether protocol work that concurrent execution becomes much higher concurrency control techniques might be. Many factors might interleave their conflict checking for ti waits for example, where to deal with greater amounts to thus in concurrency distributed system and high communication failures in mvgv and delivers it. As we tap with simple majority voting last class, Vol. It had received these protocols in concurrency control protocol is detected during the system of support acid properties which it needs a distributed.
    [Show full text]
  • An Evaluation of Distributed Concurrency Control
    An Evaluation of Distributed Concurrency Control Rachael Harding Dana Van Aken MIT CSAIL Carnegie Mellon University [email protected] [email protected] Andrew Pavlo Michael Stonebraker Carnegie Mellon University MIT CSAIL [email protected] [email protected] ABSTRACT there is little understanding of the trade-offs in a modern cloud Increasing transaction volumes have led to a resurgence of interest computing environment offering high scalability and elasticity. Few in distributed transaction processing. In particular, partitioning data of the recent publications that propose new distributed protocols across several servers can improve throughput by allowing servers compare more than one other approach. For example, none of the to process transactions in parallel. But executing transactions across papers published since 2012 in Table 1 compare against timestamp- servers limits the scalability and performance of these systems. based or multi-version protocols, and seven of them do not compare In this paper, we quantify the effects of distribution on concur- to any other serializable protocol. As a result, it is difficult to rency control protocols in a distributed environment. We evaluate six compare proposed protocols, especially as hardware and workload classic and modern protocols in an in-memory distributed database configurations vary across publications. evaluation framework called Deneva, providing an apples-to-apples Our aim is to quantify and compare existing distributed concur- comparison between each. Our results expose severe limitations of rency control protocols for in-memory DBMSs. We develop an distributed transaction processing engines. Moreover, in our anal- empirical understanding of the behavior of distributed transactions ysis, we identify several protocol-specific scalability bottlenecks.
    [Show full text]